Give your AI a route to your work.

Keep a machine at home. Continue from Claude or ChatGPT on another device. Phewsh connects project context to your AI through MCP; its outbound node can offer approved repositories and installed coding agents without exposing a public terminal port. Ion keeps the shared request, review, and record.

Remote execution preview. Phewsh's authenticated MCP connector currently shares project continuity and handoffs. The outbound node, project registration, and governed task routes are implemented; direct file browsing, arbitrary terminal commands, Git edits, PR creation, and unattended control through the provider connector are not yet generally available. Do not rely on it as a remote desktop replacement today.

Your tools stay yours. The project truth stays shared.

Your AIChatGPT · Claude
web, desktop, or mobile
Phewsh MCPauthenticated connector
continuity + tool requests
Ion + cloud relayteam permissions
review + record
Your machineoutbound node
approved repo + harness

The Mac or server initiates an outbound connection. No inbound port forwarding required. Remote execution requests still require a separate human approval.

Serve a machine for remote work

Run these on the computer holding your code, not on your phone. Each command is separate so login and verification prompts can finish before the next step. Requires Node.js and a Git repository with an origin remote. You can stop the worker at any time.

1 Install Phewsh

Install the open-source CLI. Installing alone neither registers projects nor grants remote access.

curl -fsSL phewsh.com/install.sh | sh
2 Sign in to the same Phewsh account

Complete the email verification prompt. Your terminal login must have access to the same cloud project as your connector account.

phewsh login
3 Allow one repository and link its project

Start inside the repository you want to serve. Registering it is an explicit opt-in. Use the cloud project ID for phewsh link if this repo is not already linked. Do not link unrelated repositories to the same project just to bypass a permission check.

cd /path/to/your/git-repo
phewsh project add
phewsh link <your-cloud-project-id>
phewsh ion doctor

Replace placeholders and run one line at a time. Existing correctly linked projects can skip phewsh link.

4 Start the outbound node

Start the worker after registration. Confirm Ion cloud node ... connected and a nonzero project count. Keep the computer powered, online, and awake. In foreground mode, the terminal session must remain open; automatic startup is a separate setup.

phewsh serve --cloud
5 Add Phewsh to your AI service

Use the Phewsh Dev connector where supported, or add the hosted remote MCP address as a custom connector in a compatible AI app. Follow that client's authentication flow. The endpoint is not your Mac's localhost URL.

https://mcp.phewsh.com/mcp

Current ChatGPT/Claude connector tools expose the shared project record, not full remote shell or filesystem control yet. Connector availability, permission prompts, and authentication depend on the AI service.

What works—and what's next

Available

Continuity across AI clients

Project list, active context, decisions, and handoffs through the Phewsh Dev MCP connector. Your .intent/ files stay canonical; Ion carries recorded cloud state.

Node preview

Outbound execution node

Run phewsh serve --cloud to advertise linked repos and installed coding-agent routes. Remote task contracts, human review, node leases, and receipts are being validated end-to-end.

In development

Files, terminal, Git, and PRs

Scoped remote file operations, approved shell execution, Git status and edits, creating pull requests, and merge workflows directly from remote AI clients.

Longer-term

Many people, one accountable workspace

Several authorized humans and agents sharing project-level tasks, review and receipts; not a shared unrestricted shell or a replacement for each person's credentials.

Local MCP or remote MCP?

Local adapter · available

Working inside your own computer's coding tool? Use phewsh mcp setup to configure the bundled local stdio adapter. It exposes bounded project context and doesn't make that machine reachable from mobile apps.

phewsh mcp setup

Remote connector · preview

Working through ChatGPT or Claude on the go? Connect to the hosted MCP endpoint. Separately serve a registered Mac or server as an outbound node. Remote project actions remain permissioned and reviewed; general-purpose computer tools are still being built.

phewsh serve --cloud

Why Ion is part of the picture

MCP lets your chosen AI ask for work; it isn't a second identity system or a hidden permission bypass. Ion keeps the project room, task requests, human approvals, and evidence of what actually ran. That makes the same machine useful to a team through separate authenticated identities and auditable work, rather than an undifferentiated shared desktop. The multi-person control experience is a direction under active development, not unrestricted parallel access today.

Questions before you connect

Can I leave my Mac at home and use ChatGPT or Claude on my phone?

The outbound node can stay connected while the Mac is awake and online, and the hosted connector can share Phewsh project context across devices. Direct remote file edits and terminal commands from the provider connector are still in development. Verify those tools are deployed before relying on them.

Does connecting MCP give AI unrestricted access to everything?

No. The current remote path requires a named, authorized project; a registered Git repository; an engine-bound task; and separate human approval before execution. Do not store secrets in public task descriptions or approve work you haven't reviewed.

Do I need to open Ion or forward a port on my computer?

No desktop window or inbound router port is required to run the outbound node. Ion is the review and record surface; it may still be needed for approving a remote action. The local browser diagnostic below is optional.

Can I grant access to every repo in a folder?

Per-repository identities and cloud memberships remain the security boundary. Workspace discovery is in development; it is not an implicit authorization for every subfolder, backup, worktree, or sensitive file.

Is this free?

Phewsh's open-source components can be run and self-hosted, but hosting, AI model usage, and services used to relay or run work may have separate costs. This preview does not promise unlimited free provider usage.

Optional: inspect this browser's machine

The panel below probes localhost:7483 on the computer running this browser. It cannot see your Mac at home from your phone; an offline result here does not mean your remote node is disconnected. Use the Phewsh cloud project and Ion for that.

Local bridge not connected
Connect only when you want this page to inspect the loopback worker on this machine.

Recent receipts

Redacted activity appears here when the local bridge is running. These bounded records point to what was observed; they are not transcripts or proof that every claimed result was independently verified.

How the planes fit together

/cli — command
Author intent, sync it everywhere, run the bridges. The terminal surface.
/mcp — coordination
Remote connector onboarding, local adapter, outbound nodes, permissions, and receipts.
/api — inference
One OpenAI-compatible key for any model, bounded by your project's budget.

The local browser diagnostic contacts only this browser's loopback address; it does not enumerate your other machines. The remote node initiates outbound HTTPS, with cloud membership and separate approvals for execution. Phewsh never turns your home router into an unauthenticated MCP server. See /connect for other adapters and /cli for the complete terminal guide.